Plastic deformation in crystalline materials is inherently multiscale: macroscopic stress–strain response emerges from the collective motion, interaction, and patterning of dislocations. Continuum plasticity models are computationally efficient at engineering scales but typically encode microstructural effects through phenomenological assumptions, whereas high-fidelity approaches such as discrete dislocation dynamics (DDD) resolve the governing mechanisms at high computational cost. A central barrier to predictive, scalable dislocation-based plasticity is the lack of robust closure relations that supply microstructure-dependent information required by continuum dislocation formulations from a limited set of low-order continuum fields. We present a machine-learning workflow that learns compact microstructure descriptors and latent-space evolution laws from high-fidelity dislocation simulations to enable data-driven closure and efficient surrogate modeling. DDD data are coarse-grained into continuum field representations and used to learn two complementary descriptor families: (i) low-dimensional latent variables obtained from autoencoders trained directly on coarse-grained dislocation fields, and (ii) topology-aware descriptors learned by graph neural networks (GNNs) operating on graph representations of discrete dislocation networks. By analyzing trajectories in latent space, we identify a reduced microstructural state whose temporal evolution can be modeled as a latent-dynamics problem. Using this latent evolution, we construct surrogate predictions that reproduce the main trends of the high-fidelity simulations - capturing the evolution of key microstructure-dependent quantities and associated macroscopic response with good qualitative agreement - while substantially reducing computational cost relative to DDD. Overall, the proposed approach provides a systematic pathway to transfer discrete-scale dislocation mechanisms into continuum-scale plasticity through learned descriptors and latent dynamics, advancing physics-informed machine learning for computational mechanics across scales.